Foundations of Machine Learning and its Applications for Scientific Discovery in Physical and Biological Systems Workshop

JUNE 23 – 24, 2022 | DC METRO AREA, USA


In the past 15 years, there have been tremendous developments in machine learning (ML) based on deep neural networks (DNNs). Modern artificial neural network (ANN)-based algorithms, in particular deep-learning neural networks (DLNNs), have enjoyed a long string of successes, achieving human-level performance in image recognition, machine translation, game-playing, and even solving long-standing grand-challenge scientific problems, such as protein structure prediction. However, despite their many successful applications, a systematic understanding of the underlying principles of DNNs, i.e., why they work, when they work, and how they work, is still lacking. Historically, statistical physics played an important role in the initial development of artificial neural networks, in the form of the Hopfield model, the Boltzmann machine, and applications of spin-glass theory to neural networks. We believe the time is ripe to develop a solid theoretical foundation for DNN algorithms based on concepts and methods from statistical physics, and to further develop deep-learning based applications to accelerate scientific discovery in the physical and biological world.  

The NSF Physics of Living Systems will be organizing a two-day workshop in support of bringing together disparate members of the community in furtherance of the aforementioned goals. There are two interconnected themes for this workshop. First, we aim to bring experts from the physics and machine learning community together to provide their perspectives on fundamental issues and possible directions for understanding and advancing AI research based on ideas and tools from physics.  We plan to cover a wide range of topics including theoretical foundation for learning capacity and generalization, path integral approach for reinforcement learning and its application, dynamics of stochastic-gradient-descend based learning, mean-field theory of training wide network, renormalization theory of DNN, optimal transport theory for generative modeling, loss function landscape of DNN, information bottleneck in ML, and more. Second, we aim to bring experts at the forefront of applying deep learning and AI to help improve our understanding of complex physical and biological systems. Some of the key issues we hope to address include how to incorporate physical/biochemical constraints into the architecture of the deep nets and other machine learning models, how to interpret the neural network model, and whether or not the current AI architecture/algorithms are capable of deriving/discovering laws of nature.

Herbert Levine, Northeastern University
José Onuchic, Rice University
Yuhai Tu, IBM T. J. Watson Research Center

David Baker, University of Washington via Zoom

E. Paulo Alves, UCLA
Gurinder Atwal, Regeneron Pharma
Krastan Blagoev, National Science Foundation
Pratik Chaudhari, UPENN
Cecilia Clementi, Rice University
Ethan Dyer, Google
Rafael Gomez-Bombarelli, MIT

Olexandr Isayev, Carnegie Mellon University
Dima Krotov, IBM
Steven Lopez, Northeastern University
Vyacheslav Lukin, National Science Foundation
Stephane Mallat, Collège de France
Niall Mangan, Northwestern University
Faruck Morcos, UT-Dallas
Qing Nie, UC-Irvine
Cengiz Pehlevan, Harvard University
Irina Rish, Université de Montréal
Grant Rotskoff, Stanford University
David Schwab, CUNY
James Shank, National Science Foundation
Amarda Shehu, NSF + George Mason University
Tess Smidt, MIT

Haim Sompolinsky,  The Edmond & Lily Safra Center for Brain Sciences (ELSC)
Daniel Tartakovsky, Stanford University
Robin Walters, Northeastern University
Neal Woodbury, Arizona State University